Etched, which is a four-year-old chip startup, claims to have received more than $1 billion worth of orders of hardware that is basically designed to run AI models after they have already been trained. The process of running AI models is known as inference and this has become one of the costliest operations for companies using AI in their business operations, and Etched believes that chip technology can do the job more quickly, cheaply, and with less energy use than Nvidia’s general-purpose GPUs.

The company disclosed its orders together with a valuation of $5 billion. It added that its first chip is already made by Taiwan Semiconductor Manufacturing Company (TSMC), as per a report by TechCrunch. If the chip works as expected, a portion of the money currently flowing to Nvidia would be diverted, leading to a change in the economics around deploying frontier AI models.

“We’ve hit a point in the evolution of AI where specialized chips that can perform better than general-purpose GPUs are inevitable — and the technical decision-makers of the world know this,” Etched co-founder and CEO Gavin Uberti told TechCrunch in 2024.

Unlike training where models learn from huge datasets, inference takes place whenever a user submits a prompt. Currently, large language models make use of the time looking for weights and updating their key-value (KV) cache which increases the importance of the memory capacity and bandwidth as context windows get bigger. As a result, there is a demand for chips designed for inference only as opposed to GPUs capable of handling heavier workloads.

A paper published in July 2026 by Michael J. Yuan and Ju Long states that mainstream GPUs can be characterized as “compute-heavy and capacity-light,” owing to their combination of high processing power with relatively low memory. According to the authors, during inference, some computations are often left unused while waiting for data, which creates a possibility to use hardware designed specifically for optimizing memory consumption rather than arithmetic efficiency.

What Etched is selling

Instead of marketing a standalone chip, Etched provides “frontier inference clusters” which are an assemblage of systems functioning together comprising of custom silicon, networking units and software. The firm claims that over 400 engineers are employed by it who have been recruited from Nvidia, Google TPU team, Broadcom, SK Hynix and TSMC.

The firm has attributed the expected enhancement in performance to two proprietary technologies. The first one is Low Voltage Inference (LVI), a technology meant to allow computing blocks operate at lower voltage levels to achieve continuous operation with lower heat production. The other technology is the Cluster Scale Memory (CSM), which free chips in a low-latency storage system and thus eliminates problems with long-context inference and KV-cache size growth.

These technologies, however, are still considered claims made by the company rather than being confirmed independently. Etched has claimed that it expects to make some of its benchmark results and technical information public as soon as it starts shipping its first product racks later this summer.

From near-bankruptcy to a crowded cap table

Etched has raised roughly $800 million to date, including a $500 million financing completed in December at a $5 billion post-money valuation, according to TechCrunch. Stripes led the round, with participation from Jane Street, Hudson River Trading, Two Sigma, Ribbit Capital and VentureTech Alliance.

Its investor roster also includes prominent AI researchers Andrej Karpathy, Geoffrey Hinton, Fei-Fei Li, Arthur Mensch and Scott Wu, as well as Stanley Druckenmiller and Peter Thiel.

This backing represents a complete change of direction for the startup. Etched, which was established in 2022 by Harvard dropouts and Thiel Fellows Robert Wachen and Gavin Uberti, ran into trouble trying to attract investors for most parts of 2023. The creators of the company have stated that almost every organization that they turned to dismissed their requests before they managed to get over $125 million in funding in 2024.

A widening race the industry is watching

Etched is entering an already populated market. Cerebras had one of the most highly publicized AI chip IPOs this year, Groq got a $650 million raise recently, while Amazon, Google, and Microsoft are creating their AI custom chips for their own infrastructure. OpenAI has also hired Broadcom to create a custom chip, pointing to the fact that the inference market has reached a level where it can support other players apart from Nvidia.

The manufacturing process continues to be a struggle. As reported by Cryptopolitan earlier, advanced packaging technology, especially TSMC’s CoWoS method of joining processors with high-bandwidth memory, has become one of the biggest obstacles to the industry.

The demand for AI wafers is expected to increase nearly eleven times from 2022 to 2026. Meanwhile, TSMC has about seventy-two percent of the market of the pure-play foundry. Since Etched uses TSMC’s N4P process, it ends up competing with the likes of Nvidia, AMD, and other AI chipmakers for the same manufacturing resources.

At the moment, the focus is on implementation. Reports indicate that production has already begun and that the first inference racks will be shipped in the coming months to get about $1 billion in sales orders. The extent to which hardware has delivered on the company’s promises of improved features will only be subject to independent verification once it reaches customers.

For the time being, the company’s claims about LVI and CSM should be regarded as promises rather than confirmed results. If the results turn out to be true, the start-up is likely to become one of Nvidia’s fiercest competitors in the AI inference sector.

The smartest crypto minds already read our newsletter. Want in? Join them.